2. Concept learning Maria Simi, 2011/2012 Machine Learning, Tom Mitchell Mc Graw-Hill International Editions, 1997 (Cap 1, 2).

3. Introduction to machine learning  Introduction to machine learning  When appropriate and when not appropriate  Task definition  Learning methodology: design, experiment, evaluation  Learning issues: representing hypothesis  Learning paradigms  Supervised learning  Unsupervised learning  Reinforcement learning  Next week …

4. AIMA learning architecture

5. Machine learning: definition  A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E [Mitchell]  Problem definition for a learning agent  Task T  Performance measure P  Experience E

6. Designing a learning system 1. Choosing the training experience  Examples of best moves, games outcome … 2. Choosing the target function  board-move, board-value, … 3. Choosing a representation for the target function  linear function with weights (hypothesis space) 4. Choosing a learning algorithm for approximating the target function  A method for parameter estimation

7. Design of a learning system Mitchell

8. Inductive learning  Inductive learning  Inducing a general function from training examples  A supervised paradigm  Basic schemas that assume a logical representation of the hypothesis  Concept learning  Decision trees learning  Important issues  Inductive bias (definition)  The problem of overfitting  Bibliography:  Mitchell, cap1,2,3

9. Definition of concept learning  Task: learning a category description (concept) from a set of positive and negative training examples.  Concept may be an event, an object …  Target function: a boolean function c : X  {0, 1}  Experience: a set of training instances D:{  x, c(x)  }  A search problem for best fitting hypothesis in a hypotheses space  The space is determined by the choice of representation of the hypothesis (all boolean functions or a subset)

10. Sport example  Concept to be learned: Days in which Aldo can enjoy water sport Attributes : Sky: Sunny, Cloudy, RainyWind: Strong, Weak AirTemp: Warm, ColdWater: Warm, Cool Humidity: Normal, HighForecast: Same, Change  Instances in the training set (out of the 96 possible):

11. Hypotheses representation  h is a set of constraints on attributes:  a specific value: e.g. Water = Warm  any value allowed: e.g. Water = ?  no value allowed: e.g. Water = Ø  Example hypothesis: Sky AirTemp Humidity Wind Water Forecast  Sunny,?,?,Strong, ?,Same  Corresponding to boolean function: Sunny(Sky) ∧ Strong(Wind) ∧ Same(Forecast)  H, hypotheses space, all possible h

12. Hypothesis satisfaction  An instance x satisfies an hypothesis h iff all the constraints expressed by h are satisfied by the attribute values in x.  Example 1: x 1 :  Sunny, Warm, Normal, Strong, Warm, Same  h 1 :  Sunny, ?, ?, Strong, ?, Same  Satisfies? Yes  Example 2: x 2 :  Sunny, Warm, Normal, Strong, Warm, Same  h 2 :  Sunny, ?, ?, Ø, ?, Same  Satisfies? No

13. Formal task description  Given:  X all possible days, as described by the attributes  A set of hypothesis H, a conjunction of constraints on the attributes, representing a function h: X  {0, 1} [ h ( x ) = 1 if x satisfies h ; h ( x ) = 0 if x does not satisfy h]  A target concept: c: X  {0, 1} where c(x) = 1 iff EnjoySport = Yes ; c(x) = 0 iff EnjoySport = No ;  A training set of possible instances D : {  x, c(x)  }  Goal: find a hypothesis h in H such that h(x) = c(x) for all x in D Hopefully h will be able to predict outside D…

14. The inductive learning assumption  We can at best guarantee that the output hypothesis fits the target concept over the training data  Assumption: an hypothesis that approximates well the training data will also approximate the target function over unobserved examples  i.e. given a significant training set, the output hypothesis is able to make predictions

15. Concept learning as search  Concept learning is a task of searching an hypotheses space  The representation chosen for hypotheses determines the search space  In the example we have:  3 x 2 5 = 96 possible instances (6 attributes)  1 + 4 x 3 5 = 973 possible hypothesis considering that all the hypothesis with some  are semantically equivalent, i.e. inconsistent  Structuring the search space may help in searching more efficiently

16. General to specific ordering  Consider: h 1 =  Sunny, ?, ?, Strong, ?, ?  h 2 =  Sunny, ?, ?, ?, ?, ?   Any instance classified positive by h 1 will also be classified positive by h 2  h 2 is more general than h 1  Definition: h j  g h k iff (  x  X ) [(h k = 1)  (h j = 1)]  g more general or equal; > g strictly more general  Most general hypothesis:  ?, ?, ?, ?, ?, ?   Most specific hypothesis:  Ø, Ø, Ø, Ø, Ø, Ø 

17. General to specific ordering: induced structure

18. Find-S : finding the most specific hypothesis  Exploiting the structure we have alternatives to enumeration … 1. Initialize h to the most specific hypothesis in H 2. For each positive training instance: for each attribute constraint a in h: If the constraint a is satisfied by x then do nothing else replace a in h by the next more general constraint satified by x (move towards a more general hp) 3. Output hypothesis h

19. Find-S in action

20. Properties of Find-S  Find-S is guaranteed to output the most specific hypothesis within H that is consistent with the positive training examples  The final hypothesis will also be consistent with the negative examples  Problems:  There can be more than one “most specific hypotheses”  We cannot say if the learner converged to the correct target  Why choose the most specific?  If the training examples are inconsistent, the algorithm can be mislead: no tolerance to rumor.  Negative example are not considered

21. Candidate elimination algorithm: the idea  The idea: output a description of the set of all hypotheses consistent with the training examples (correctly classify training examples).  Version space: a representation of the set of hypotheses which are consistent with D 1. an explicit list of hypotheses (List-Than-Eliminate) 2. a compact representation of hypotheses which exploits the more_general_than partial ordering (Candidate- Elimination)

22. Version space  The version space VS H,D is the subset of the hypothesis from H consistent with the training example in D VS H,D  {h  H | Consistent(h, D)}  An hypothesis h is consistent with a set of training examples D iff h ( x ) = c ( x ) for each example in D Consistent(h, D)  (   x, c(x)   D) h(x) = c(x)) Note : " x satisfies h " ( h ( x )= 1 ) different from “h consistent with x" In particular when an hypothesis h is consistent with a negative example d =  x, c( x )=No , then x must not satisfy h

23. The List-Then-Eliminate algorithm Version space as list of hypotheses 1. V ersionSpace  a list containing every hypothesis in H 2. For each training example,  x, c(x)  Remove from VersionSpace any hypothesis h for which h(x)  c(x) 3. Output the list of hypotheses in VersionSpace  Problems  The hypothesis space must be finite  Enumeration of all the hypothesis, rather inefficient

24. A compact representation for Version Space  Version space represented by its most general members G and its most specific members S (boundaries) Note: The output of Find-S is just  Sunny, Warm, ?, Strong, ?, ? 

25. General and specific boundaries  The Specific boundary, S, of version space VS H,D is the set of its minimally general (most specific) members S  {s  H | Consistent(s, D)  (  s'  H)[(s  g s')  Consistent(s', D)]} Note: any member of S is satisfied by all positive examples, but more specific hypotheses fail to capture some  The General boundary, G, of version space VS H,D is the set of its maximally general members G  {g  H | Consistent(g, D)  (  g'  H)[(g'  g g)  Consistent(g', D)]} Note: any member of G is satisfied by no negative example but more general hypothesis cover some negative example

26. Version Space representation theorem  G and S completely define the Version Space  Theorem: Every member of the version space ( h consistent with D ) is in S or G or lies between these boundaries VS H,D ={h  H |(  s  S) (  g  G) (g  g h  g s)} where x  g y means x is more general or equal to y Sketch of proof:  If g  g h  g s, since s is in S and h  g s, h is satisfied by all positive examples in D; g is in G and g  g h, then h is satisfied by no negative examples in D; therefore h belongs to VS H,D  It can be proved by assuming a consistent h that does not satisfy the right-hand side and by showing that this would lead to a contradiction

27. Candidate elimination algorithm-1 S  minimally general hypotheses in H, G  maximally general hypotheses in H Initially any hypothesis is still possible S 0 = , , , , ,  G 0 =  ?, ?, ?, ?, ?, ?  For each training example d, do: If d is a positive example: 1. Remove from G any h inconsistent with d 2. Generalize(S, d) If d is a negative example: 1. Remove from S any h inconsistent with d 2. Specialize ( G, d ) Note: when d =  x, No  is a negative example, an hypothesis h is inconsistent with d iff h satisfies x

28. Candidate elimination algorithm-2 Generalize(S, d): d is positive For each hypothesis s in S not consistent with d : 1. Remove s from S 2. Add to S all minimal generalizations of s consistent with d and having a generalization in G 3. Remove from S any hypothesis with a more specific h in S Specialize(G, d): d is negative For each hypothesis g in G not consistent with d : i.e. g satisfies d, 1. Remove g from G but d is negative 2. Add to G all minimal specializations of g consistent with d and having a specialization in S 3. Remove from G any hypothesis having a more general hypothesis in G

29. Example: initially , , , , .  S0:S0:  ?, ?, ?, ?, ?, ?  G0G0

30. Example: after seing  Sunny,Warm, Normal, Strong, Warm, Same  +  Sunny,Warm, Normal, Strong, Warm, Same  S1:S1: , , , , .  S0:S0:  ?, ?, ?, ?, ?, ?  G 0, G 1

31. Example: after seing  Sunny,Warm, High, Strong, Warm, Same  +  Sunny,Warm, Normal, Strong, Warm, Same  S1:S1:  ?, ?, ?, ?, ?, ?  G 1, G 2  Sunny,Warm, ?, Strong, Warm, Same  S2:S2:

32. Example: after seing  Rainy, Cold, High, Strong, Warm, Change  S 2, S 3 :  ?, ?, ?, ?, ?, ?  G2:G2:  Sunny, Warm, ?, Strong, Warm, Same   Sunny, ?, ?, ?, ?, ?   ?, Warm, ?, ?, ?, ?   ?, ?, ?, ?, ?, Same  G3:G3:

33. Example: after seing  Sunny, Warm, High, Strong, Cool Change  + S3S3 G3:G3:  Sunny, Warm, ?, Strong, Warm, Same   Sunny, ?, ?, ?, ?, ?   ?, Warm, ?, ?, ?, ?   ?, ?, ?, ?, ?, Same  G4:G4:  Sunny, Warm, ?, Strong, ?, ?  S4S4  Sunny, ?, ?, ?, ?, ?   ?, Warm, ?, ?, ?, ? 

34. Learned Version Space

35. Observations  The learned Version Space correctly describes the target concept, provided: 1. There are no errors in the training examples 2. There is some hypothesis that correctly describes the target concept  If S and G converge to a single hypothesis the concept is exactly learned  In case of errors in the training, useful hypothesis are discarded, no recovery possible  An empty version space means no hypothesis in H is consistent with training examples

36. Ordering on training examples  The learned version space does not change with different orderings of training examples  Efficiency does  Optimal strategy (if you are allowed to choose)  Generate instances that satisfy half the hypotheses in the current version space. For example:  Sunny, Warm, Normal, Light, Warm, Same  satisfies 3/6 hyp.  Ideally the VS can be reduced by half at each experiment  Correct target found in  log 2 |VS|  experiments

37. Use of partially learned concepts Classified as positive by all hypothesis, since satisfies any hypothesis in S

38. Classifying new examples Classified as negative by all hypothesis, since does not satisfy any hypothesis in G

39. Classifying new examples Uncertain classification: half hypothesis are consistent, half are not consistent

40. Classifying new examples  Sunny, Cold, Normal, Strong, Warm, Same  4 hypothesis not satisfied; 2 satisfied Probably a negative instance. Majority vote?

41. Hypothesis space and bias  What if H does not contain the target concept?  Can we improve the situation by extending the hypothesis space?  Will this influence the ability to generalize?  These are general questions for inductive inference, addressed in the context of Candidate-Elimination  Suppose we include in H every possible hypothesis … including the ability to represent disjunctive concepts

42. Extending the hypothesis space  No hypothesis consistent with the three examples with the assumption that the target is a conjunction of constraints  ?, Warm, Normal, Strong, Cool, Change  is too general  Target concept exists in a different space H', including disjunction and in particular the hypothesis Sky=Sunny or Sky=Cloudy SkyAirTempHumidityWindWaterForecastEnjoyS 1SunnyWarmNormalStrongCoolChangeYES 2CloudyWarmNormalStrongCoolChangeYES 3RainyWarmNormalStrongCoolChangeNO

43. An unbiased learner  Every possible subset of X is a possible target |H'| = 2 |X|, or 2 96 (vs |H| = 973, a strong bias)  This amounts to allowing conjunction, disjunction and negation  Sunny, ?, ?, ?, ?, ?  V <Cloudy, ?, ?, ?, ?, ?  Sunny(Sky) V Cloudy(Sky)  We are guaranteed that the target concept exists  No generalization is however possible!!! Let's see why …

44. No generalization without bias!  VS after presenting three positive instances x 1, x 2, x 3, and two negative instances x 4, x 5 S = {( x 1 v x 2 v x 3 )} G = {¬( x 4 v x 5 )} …all subsets including x 1 x 2 x 3 and not including x 4 x 5  We can only classify precisely examples already seen!  Take a majority vote?  Unseen instances, e.g. x, are classified positive (and negative) by half of the hypothesis  For any hypothesis h that classifies x as positive, there is a complementary hypothesis ¬ h that classifies x as negative

45. No inductive inference without a bias  A learner that makes no a priori assumptions regarding the identity of the target concept, has no rational basis for classifying unseen instances  The inductive bias of a learner are the assumptions that justify its inductive conclusions or the policy adopted for generalization  Different learners can be characterized by their bias  See next for a more formal definition of inductive bias …

46. Inductive bias: definition  Given:  a concept learning algorithm L for a set of instances X  a concept c defined over X  a set of training examples for c: D c = {  x, c(x)  }  L(x i, D c ) outcome of classification of x i after learning  Inductive inference ( ≻ ): D c  x i ≻ L(x i, D c )  The inductive bias is defined as a minimal set of assumptions B, such that (| − for deduction)  (x i  X) [ (B  D c  x i ) | − L(x i, D c ) ]

47. Inductive bias of Candidate-Elimination  Assume L is defined as follows:  compute VS H,D  classify new instance by complete agreement of all the hypotheses in VS H,D  Then the inductive bias of Candidate-Elimination is simply B  (c  H)  In fact by assuming c  H: 1. c  VS H,D, in fact VS H,D includes all hypotheses in H consistent with D 2. L(x i, D c ) outputs a classification "by complete agreement", hence any hypothesis, including c, outputs L(x i, D c )

48. Inductive system

49. Equivalent deductive system

50. Each learner has an inductive bias  Three learner with three different inductive bias: 1. Rote learner: no inductive bias, just stores examples and is able to classify only previously observed examples 2. CandidateElimination: the concept is a conjunction of constraints. 3. Find-S: the concept is in H (a conjunction of constraints) plus "all instances are negative unless seen as positive examples” (stronger bias)  The stronger the bias, greater the ability to generalize and classify new instances (greater inductive leaps).

51. Bibliography  Machine Learning, Tom Mitchell, Mc Graw-Hill International Editions, 1997 (Cap 2).